Table 2: Performance of representative models from the recent literature, and of the analogous model that we use in this study.

Both KronRLS and SimBoost use Smith-Waterman and PubChem similarity matrices to featurize proteins and inhibitors, whereas DeepDTA uses a CNN. Performance is reported as the concordance index (CI) and the mean square error (MSE); a model with good performance should have high concordance index (CI) and low mean square error (MSE).

Model	Dataset	Protein Representation	Ligand Representation	CI	MSE
KronRLS	KIBA	S-W ➔ Dense Network	PubChem ➔ Dense Network	0.782	0.441
SimBoost	KIBA	S-W ➔ Dense Network	PubChem ➔ Dense Network	0.836	0.222
DeepDTA	KIBA	CNN	CNN	0.863	0.194
KronRLS	Davis	S-W ➔ Dense Network	PubChem ➔ Dense Network	0.782	0.379
SimBoost	Davis	S-W ➔ Dense Network	PubChem ➔ Dense Network	0.872	0.282
DeepDTA	Davis	CNN	CNN	0.878	0.261
Our model	Davis	CNN	CNN	0.896	0.177